A developer shared a detailed cost breakdown for using LLM API relays, highlighting significant savings achieved through prompt caching and group rate multipliers. The author demonstrated how prompt cache hits can reduce costs by up to 87.5%, especially in agentic workflows that repeatedly send large prompts. Additionally, using tiered rate multipliers for different model groups, such as open-weight models at 0.08x list price, further compounded savings, leading to an overall reduction of 92% against list prices. The post also provides practical advice for users to sanity-check relay services, including verifying model lists, checking for cache hit/miss data, and comparing responses against official APIs, while cautioning about the lack of enterprise SLAs and potential instability of cheaper relays. AI
IMPACT Provides practical strategies for optimizing LLM API costs, potentially influencing how developers manage their AI infrastructure expenses.
RANK_REASON Developer shares personal cost breakdown and tips for using LLM API relays, not a primary release or industry-shaking event.
- api.dshapi.icu
- Claude Code
- Codex Cli
- DeepSeek-V4 Flash
- DeepSeek V4-Pro
- GLM-5.2
- GLM 5.3
- HY4
- Kimi k3
- Minimax M3
- tencent/Hy3
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
